What changed
Inception announced Mercury 3 on October 19, 2026 in a blog post as its new flagship model. Unlike most chat models, which write one token after another, Mercury is a diffusion language model: it drafts a whole response and refines many positions in parallel, which Inception says is the source of its speed.
According to Inception, Mercury 3 is more than 40 times larger than Mercury 1 (the exact parameter count is not disclosed), raises the context window to 1 million tokens from 260,000 in Mercury 2.5, and generates about 1,068 tokens per second on NVIDIA GPUs, which it presents as 4 to 8 times faster than its rivals. In the company's comparison (all at medium reasoning settings), GPT-6 Luna reached about 130 and Claude Haiku 5.5 about 240 tokens per second.
Who can use it and when
Mercury 3 is available now through the Inception API Platform and Mercury Chat, and through OpenRouter, Baseten and Vercel. The API is OpenAI-compatible, so existing code that calls an OpenAI-style chat endpoint should need only a base URL, key and model name change. OpenRouter lists it as inception/mercury-3 with a 1,000,000-token context, up to 131,072 output tokens, and support for tool calling and structured outputs.
On OpenRouter, Inception is currently the only provider, and the model is not available through EU in-region routing, which matters for teams with EU data-residency requirements. Inception says new API users get 100 million free tokens to try it. Enterprise deployments with dedicated capacity and configurable data retention are offered through sales.
Why it matters
Speed matters most where a model is called many times in a row or a person is waiting: browser agents that take dozens of steps, voice and chat support, real-time search and tutoring. Early-access customers quoted by Inception report answers in about a second; search startup Octen AI cited a median generation time of 0.9 seconds.
On quality, Inception's own table (all three models at medium reasoning settings) shows Mercury 3 ahead of both rivals on 8 of 12 benchmarks, with its largest lead over GPT-6 Luna in agentic coding (51.08 on SWE-Bench Atlas versus 33.33). It trails Claude Haiku 5.5 on DeepSWE and MRCR v2, and GPT-6 Luna on AA-LCR and AA-Omniscience.
Access and limitations
List pricing is $0.20 per million input tokens and $0.70 per million output tokens, with cached input 90% cheaper. A launch promotion halves this to $0.10 and $0.35; Inception has not said when the promotion ends. GPT-6 Luna and Claude Haiku 5.5 each list at $0.10 input and $0.50 output for typical prompts, according to SiliconANGLE. At list price Mercury 3 costs more than both on input ($0.20 vs $0.10) and output ($0.70 vs $0.50); only during the launch promotion is it cheaper on output, while input then matches them.
The benchmark and speed figures come from Inception and its customers, and no outside group has published its own benchmark results yet. OpenRouter's live stats, checked on October 21, 2026, showed a median of about 240 to 250 tokens per second and a median latency of about 0.9 seconds, well below Inception's 1,068 figure, which was measured under the company's own conditions. Real-world speed will depend on provider, prompt length and settings. AppFor.ai has not tested the model.
Go to the source
Sources & further reading
Based on the official documentation below. Features and access can change; check the provider before making a decision.
Was this page useful?
Help us make our coverage more useful. This is feedback on the page, not a rating of the tool.